A new era of discovery: AI and the frontiers of science with Demis Hassabis

The guest in this episode is Demis Hassabis, founder of DeepMind. Instead of revisiting the well-worn AlphaFold story, he goes much further—offering a hard timeline: by around 2027, we will see AI systems that can consistently perform PhD-level research tasks. This is not a semantic debate about AGI definitions. It’s a pragmatic call: once machines can reliably complete the kind of work that takes years of doctoral training, the world will have already changed. Hassabis anchors the threshold for “powerful AI” not to a fuzzy AGI concept but to that of a “Nobel-laureate-level scientist”—and that bar is now only a few years away.

His real bet is that alignment and scaling must move in lockstep. The comfortable idea that safety research can be postponed until AI gets stronger is, in his view, completely wrong. Safety is not an afterthought—it’s part of the process itself. While chips and compute are physical bottlenecks, he argues they will not be the binding constraint before 2027; the real bottlenecks lie in data quality, algorithmic efficiency, and organizational capacity. Labs that merely pile on parameters without weaving safety into the model’s DNA will end up with systems that are powerful yet uncontrollable. This is why his push for export controls is not about blocking technology—it’s about buying time for alignment research, a stance that feels both pragmatic and urgent.

To make this concrete, he points to industry-wide practices: techniques like reinforcement learning from human feedback (RLHF) and Constitutional AI are essentially shifting value oversight from post-hoc review into the training process itself. The same philosophy shapes the business model. While many teams rush to consumer-facing apps, Hassabis stands for an API-first path—willing to cede the to-consumer innovation layer to partners, while doubling down on delivering robust agentic capabilities at the base. It sounds restrained, even risky, but the logic is clear: if the underlying model’s reliability and safety curve haven’t been proven, any flashy application built on top is just a castle on sand.

Naturally, this approach invites skepticism. Some argue that API-first will miss early consumer-market advantages, and that layering on safety processes may slow the scaling race—especially when there are rivals hungry for the “first” title. But Hassabis’s confidence comes from a single-minded belief: he’s not racing to win the AGI competition; he’s racing to get the safety curve right first. Throughout the interview, he reminds the audience that when AI begins to tackle the hard core of scientific discovery, the margin for error is far narrower than in any consumer product.

The episode covers everything from opening remarks and the origins of agent concepts, to model selection, commercialization paths, a five-year outlook, and live Q&A—a dense, high-speed journey. If you once thought AlphaFold was the end of the story, this conversation shows it may well be just the beginning. For Hassabis, all attention is now focused on a larger question: how to bring the smartest machines into the next era of scientific discovery in the most responsible way.